Synthetic demo dataHuman review requiredRuns entirely in your browser

HAL Campus Evidence Desk

Evidence synthesis for campus operations where every material claim must cite supplied evidence. Model drafts pass through a deterministic acceptance layer before a human reviewer sees them. This page runs a browser port of that layer, tested against the Python engine for identical results.

Try the acceptance gate

Pick a model draft or edit the JSON, then run it through the gate. Findings and actions survive only if they cite evidence IDs from the selected case; conflicts survive only if they match the case's seeded conflict.

Run a draft to see what the gate accepts and rejects.

What is verified

30,000reports in the committed deterministic benchmark, 0 acceptance-invariant failures
3/3canonical synthetic cases validated
25drafts and edge cases where this page and the Python engine must agree exactly
4CI platforms: Ubuntu x64 and ARM64, Windows, macOS ARM64

Limits

All data is synthetic. Conflict relations are checked against each case's seeded ground truth; a real deployment needs institution-reviewed conflict rules. Live model inference is disabled on this public page to avoid uncontrolled external compute; the separately recorded live-model receipt is in the repository. Recommendations are advisory and always end at human review.